1 /*
2 * Copyright © 2012 Rob Clark <[email protected]>
3 * SPDX-License-Identifier: MIT
4 *
5 * Authors:
6 * Rob Clark <[email protected]>
7 * Jonathan Marek <[email protected]>
8 */
9
10 #include "nir/tgsi_to_nir.h"
11 #include "pipe/p_state.h"
12 #include "tgsi/tgsi_dump.h"
13 #include "util/format/u_format.h"
14 #include "util/u_inlines.h"
15 #include "util/u_memory.h"
16 #include "util/u_string.h"
17
18 #include "freedreno_program.h"
19
20 #include "ir2/instr-a2xx.h"
21 #include "fd2_program.h"
22 #include "fd2_texture.h"
23 #include "fd2_util.h"
24 #include "ir2.h"
25
26 static struct fd2_shader_stateobj *
create_shader(struct pipe_context * pctx,gl_shader_stage type)27 create_shader(struct pipe_context *pctx, gl_shader_stage type)
28 {
29 struct fd2_shader_stateobj *so = CALLOC_STRUCT(fd2_shader_stateobj);
30 if (!so)
31 return NULL;
32 so->type = type;
33 so->is_a20x = is_a20x(fd_context(pctx)->screen);
34 return so;
35 }
36
37 static void
delete_shader(struct fd2_shader_stateobj * so)38 delete_shader(struct fd2_shader_stateobj *so)
39 {
40 if (!so)
41 return;
42 ralloc_free(so->nir);
43 for (int i = 0; i < ARRAY_SIZE(so->variant); i++)
44 free(so->variant[i].info.dwords);
45 free(so);
46 }
47
48 static void
emit(struct fd_ringbuffer * ring,gl_shader_stage type,struct ir2_shader_info * info,struct util_dynarray * patches)49 emit(struct fd_ringbuffer *ring, gl_shader_stage type,
50 struct ir2_shader_info *info, struct util_dynarray *patches)
51 {
52 unsigned i;
53
54 assert(info->sizedwords);
55
56 OUT_PKT3(ring, CP_IM_LOAD_IMMEDIATE, 2 + info->sizedwords);
57 OUT_RING(ring, type == MESA_SHADER_FRAGMENT);
58 OUT_RING(ring, info->sizedwords);
59
60 if (patches)
61 util_dynarray_append(patches, uint32_t *,
62 &ring->cur[info->mem_export_ptr]);
63
64 for (i = 0; i < info->sizedwords; i++)
65 OUT_RING(ring, info->dwords[i]);
66 }
67
68 static int
ir2_glsl_type_size(const struct glsl_type * type,bool bindless)69 ir2_glsl_type_size(const struct glsl_type *type, bool bindless)
70 {
71 return glsl_count_attribute_slots(type, false);
72 }
73
74 static void *
fd2_fp_state_create(struct pipe_context * pctx,const struct pipe_shader_state * cso)75 fd2_fp_state_create(struct pipe_context *pctx,
76 const struct pipe_shader_state *cso)
77 {
78 struct fd2_shader_stateobj *so = create_shader(pctx, MESA_SHADER_FRAGMENT);
79 if (!so)
80 return NULL;
81
82 so->nir = (cso->type == PIPE_SHADER_IR_NIR)
83 ? cso->ir.nir
84 : tgsi_to_nir(cso->tokens, pctx->screen, false);
85
86 NIR_PASS_V(so->nir, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
87 ir2_glsl_type_size, (nir_lower_io_options)0);
88
89 if (ir2_optimize_nir(so->nir, true))
90 goto fail;
91
92 so->first_immediate = so->nir->num_uniforms;
93
94 ir2_compile(so, 0, NULL);
95
96 ralloc_free(so->nir);
97 so->nir = NULL;
98 return so;
99
100 fail:
101 delete_shader(so);
102 return NULL;
103 }
104
105 static void
fd2_fp_state_delete(struct pipe_context * pctx,void * hwcso)106 fd2_fp_state_delete(struct pipe_context *pctx, void *hwcso)
107 {
108 struct fd2_shader_stateobj *so = hwcso;
109 delete_shader(so);
110 }
111
112 static void *
fd2_vp_state_create(struct pipe_context * pctx,const struct pipe_shader_state * cso)113 fd2_vp_state_create(struct pipe_context *pctx,
114 const struct pipe_shader_state *cso)
115 {
116 struct fd2_shader_stateobj *so = create_shader(pctx, MESA_SHADER_VERTEX);
117 if (!so)
118 return NULL;
119
120 so->nir = (cso->type == PIPE_SHADER_IR_NIR)
121 ? cso->ir.nir
122 : tgsi_to_nir(cso->tokens, pctx->screen, false);
123
124 NIR_PASS_V(so->nir, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
125 ir2_glsl_type_size, (nir_lower_io_options)0);
126
127 if (ir2_optimize_nir(so->nir, true))
128 goto fail;
129
130 so->first_immediate = so->nir->num_uniforms;
131
132 /* compile binning variant now */
133 ir2_compile(so, 0, NULL);
134
135 return so;
136
137 fail:
138 delete_shader(so);
139 return NULL;
140 }
141
142 static void
fd2_vp_state_delete(struct pipe_context * pctx,void * hwcso)143 fd2_vp_state_delete(struct pipe_context *pctx, void *hwcso)
144 {
145 struct fd2_shader_stateobj *so = hwcso;
146 delete_shader(so);
147 }
148
149 static void
patch_vtx_fetch(struct fd_context * ctx,struct pipe_vertex_element * elem,instr_fetch_vtx_t * instr,uint16_t dst_swiz)150 patch_vtx_fetch(struct fd_context *ctx, struct pipe_vertex_element *elem,
151 instr_fetch_vtx_t *instr, uint16_t dst_swiz) assert_dt
152 {
153 struct surface_format fmt = fd2_pipe2surface(elem->src_format);
154
155 instr->dst_swiz = fd2_vtx_swiz(elem->src_format, dst_swiz);
156 instr->format_comp_all = fmt.sign == SQ_TEX_SIGN_SIGNED;
157 instr->num_format_all = fmt.num_format;
158 instr->format = fmt.format;
159 instr->exp_adjust_all = fmt.exp_adjust;
160 instr->stride = elem->src_stride;
161 instr->offset = elem->src_offset;
162 }
163
164 static void
patch_fetches(struct fd_context * ctx,struct ir2_shader_info * info,struct fd_vertex_stateobj * vtx,struct fd_texture_stateobj * tex)165 patch_fetches(struct fd_context *ctx, struct ir2_shader_info *info,
166 struct fd_vertex_stateobj *vtx,
167 struct fd_texture_stateobj *tex) assert_dt
168 {
169 for (int i = 0; i < info->num_fetch_instrs; i++) {
170 struct ir2_fetch_info *fi = &info->fetch_info[i];
171
172 instr_fetch_t *instr = (instr_fetch_t *)&info->dwords[fi->offset];
173 if (instr->opc == VTX_FETCH) {
174 unsigned idx =
175 (instr->vtx.const_index - 20) * 3 + instr->vtx.const_index_sel;
176 patch_vtx_fetch(ctx, &vtx->pipe[idx], &instr->vtx, fi->vtx.dst_swiz);
177 continue;
178 }
179
180 assert(instr->opc == TEX_FETCH);
181 instr->tex.const_idx = fd2_get_const_idx(ctx, tex, fi->tex.samp_id);
182 instr->tex.src_swiz = fi->tex.src_swiz;
183 }
184 }
185
186 void
fd2_program_emit(struct fd_context * ctx,struct fd_ringbuffer * ring,struct fd_program_stateobj * prog)187 fd2_program_emit(struct fd_context *ctx, struct fd_ringbuffer *ring,
188 struct fd_program_stateobj *prog)
189 {
190 struct fd2_shader_stateobj *fp = NULL, *vp;
191 struct ir2_shader_info *fpi, *vpi;
192 struct ir2_frag_linkage *f;
193 uint8_t vs_gprs, fs_gprs = 0, vs_export = 0;
194 enum a2xx_sq_ps_vtx_mode mode = POSITION_1_VECTOR;
195 bool binning = (ctx->batch && ring == ctx->batch->binning);
196 unsigned variant = 0;
197
198 vp = prog->vs;
199
200 /* find variant matching the linked fragment shader */
201 if (!binning) {
202 fp = prog->fs;
203 for (variant = 1; variant < ARRAY_SIZE(vp->variant); variant++) {
204 /* if checked all variants, compile a new variant */
205 if (!vp->variant[variant].info.sizedwords) {
206 ir2_compile(vp, variant, fp);
207 break;
208 }
209
210 /* check if fragment shader linkage matches */
211 if (!memcmp(&vp->variant[variant].f, &fp->variant[0].f,
212 sizeof(struct ir2_frag_linkage)))
213 break;
214 }
215 assert(variant < ARRAY_SIZE(vp->variant));
216 }
217
218 vpi = &vp->variant[variant].info;
219 fpi = &fp->variant[0].info;
220 f = &fp->variant[0].f;
221
222 /* clear/gmem2mem/mem2gmem need to be changed to remove this condition */
223 if (prog != &ctx->solid_prog && prog != &ctx->blit_prog[0]) {
224 patch_fetches(ctx, vpi, ctx->vtx.vtx, &ctx->tex[PIPE_SHADER_VERTEX]);
225 if (fp)
226 patch_fetches(ctx, fpi, NULL, &ctx->tex[PIPE_SHADER_FRAGMENT]);
227 }
228
229 emit(ring, MESA_SHADER_VERTEX, vpi,
230 binning ? &ctx->batch->shader_patches : NULL);
231
232 if (fp) {
233 emit(ring, MESA_SHADER_FRAGMENT, fpi, NULL);
234 fs_gprs = (fpi->max_reg < 0) ? 0x80 : fpi->max_reg;
235 vs_export = MAX2(1, f->inputs_count) - 1;
236 }
237
238 vs_gprs = (vpi->max_reg < 0) ? 0x80 : vpi->max_reg;
239
240 if (vp->writes_psize && !binning)
241 mode = POSITION_2_VECTORS_SPRITE;
242
243 /* set register to use for param (fragcoord/pointcoord/frontfacing) */
244 OUT_PKT3(ring, CP_SET_CONSTANT, 2);
245 OUT_RING(ring, CP_REG(REG_A2XX_SQ_CONTEXT_MISC));
246 OUT_RING(ring,
247 A2XX_SQ_CONTEXT_MISC_SC_SAMPLE_CNTL(CENTERS_ONLY) |
248 COND(fp, A2XX_SQ_CONTEXT_MISC_PARAM_GEN_POS(f->inputs_count)) |
249 /* we need SCREEN_XY for both fragcoord and frontfacing */
250 A2XX_SQ_CONTEXT_MISC_SC_OUTPUT_SCREEN_XY);
251
252 OUT_PKT3(ring, CP_SET_CONSTANT, 2);
253 OUT_RING(ring, CP_REG(REG_A2XX_SQ_PROGRAM_CNTL));
254 OUT_RING(ring,
255 A2XX_SQ_PROGRAM_CNTL_PS_EXPORT_MODE(2) |
256 A2XX_SQ_PROGRAM_CNTL_VS_EXPORT_MODE(mode) |
257 A2XX_SQ_PROGRAM_CNTL_VS_RESOURCE |
258 A2XX_SQ_PROGRAM_CNTL_PS_RESOURCE |
259 A2XX_SQ_PROGRAM_CNTL_VS_EXPORT_COUNT(vs_export) |
260 A2XX_SQ_PROGRAM_CNTL_PS_REGS(fs_gprs) |
261 A2XX_SQ_PROGRAM_CNTL_VS_REGS(vs_gprs) |
262 COND(fp && fp->need_param, A2XX_SQ_PROGRAM_CNTL_PARAM_GEN) |
263 COND(!fp, A2XX_SQ_PROGRAM_CNTL_GEN_INDEX_VTX));
264 }
265
266 void
fd2_prog_init(struct pipe_context * pctx)267 fd2_prog_init(struct pipe_context *pctx)
268 {
269 struct fd_context *ctx = fd_context(pctx);
270 struct fd_program_stateobj *prog;
271 struct fd2_shader_stateobj *so;
272 struct ir2_shader_info *info;
273 instr_fetch_vtx_t *instr;
274
275 pctx->create_fs_state = fd2_fp_state_create;
276 pctx->delete_fs_state = fd2_fp_state_delete;
277
278 pctx->create_vs_state = fd2_vp_state_create;
279 pctx->delete_vs_state = fd2_vp_state_delete;
280
281 fd_prog_init(pctx);
282
283 /* XXX maybe its possible to reuse patch_vtx_fetch somehow? */
284
285 prog = &ctx->solid_prog;
286 so = prog->vs;
287 ir2_compile(prog->vs, 1, prog->fs);
288
289 #define IR2_FETCH_SWIZ_XY01 0xb08
290 #define IR2_FETCH_SWIZ_XYZ1 0xa88
291
292 info = &so->variant[1].info;
293
294 instr = (instr_fetch_vtx_t *)&info->dwords[info->fetch_info[0].offset];
295 instr->const_index = 26;
296 instr->const_index_sel = 0;
297 instr->format = FMT_32_32_32_FLOAT;
298 instr->format_comp_all = false;
299 instr->stride = 12;
300 instr->num_format_all = true;
301 instr->dst_swiz = IR2_FETCH_SWIZ_XYZ1;
302
303 prog = &ctx->blit_prog[0];
304 so = prog->vs;
305 ir2_compile(prog->vs, 1, prog->fs);
306
307 info = &so->variant[1].info;
308
309 instr = (instr_fetch_vtx_t *)&info->dwords[info->fetch_info[0].offset];
310 instr->const_index = 26;
311 instr->const_index_sel = 1;
312 instr->format = FMT_32_32_FLOAT;
313 instr->format_comp_all = false;
314 instr->stride = 8;
315 instr->num_format_all = false;
316 instr->dst_swiz = IR2_FETCH_SWIZ_XY01;
317
318 instr = (instr_fetch_vtx_t *)&info->dwords[info->fetch_info[1].offset];
319 instr->const_index = 26;
320 instr->const_index_sel = 0;
321 instr->format = FMT_32_32_32_FLOAT;
322 instr->format_comp_all = false;
323 instr->stride = 12;
324 instr->num_format_all = false;
325 instr->dst_swiz = IR2_FETCH_SWIZ_XYZ1;
326 }
327